theaniketmaurya

First grok cli and now this

that's why you need to use a sandbox.

SmolVM can easily run Pi, codex and claude https://github.com/CelestoAI/smolVM

Celesto has an unified interface for sandboxes based on QEMU, Firecracker and Libkrun.

https://docs.celesto.ai/smolvm

This test creates a 10 GiB sandbox, then writes 20 GiB of data into the workspace. It shows that large workspace data is not constrained by the root disk size.

1. Create a new sandbox with 10gb root storage:

`celesto computer create --template coding-agent --disk-size-mb 10240`

2. Write 20 GiB under user home dir:

``` celesto computer run einstein "python3 - <<'PY' from pathlib import Path import hashlib import os

target_dir = Path('/home/ohm/storage-proof') target_dir.mkdir(parents=True, exist_ok=True) target = target_dir / 'twenty-gib.bin'

with target.open('wb') as f: for gib in range(20): for block_index in range(1024): seed = f'{gib}:{block_index}'.encode() block = hashlib.sha256(seed).digest() * 32768 f.write(block) f.flush() os.fsync(f.fileno()) print(f'wrote {gib + 1} GiB', flush=True)

print('final size bytes:', target.stat().st_size) PY" ```

3. Verify the file is larger than the 10 GiB root disk:

`celesto computer run einstein "du -h /home/ohm/storage-proof/twenty-gib.bin && df -h / /home/ohm"`

SmolFS has a distributed posix filesystem. we use juicefs which separates the actual data and metadata (permissions, file size changes etc). Metadata store provides atomicity and locking primitives which keeps the file system horizontally scalable while maintaining concurrent mounts and handle concurrent agent operations.

Yeah, it’s quite handy that I can mount the same storage on my macbook and then somewhere on a random cloud.

Nice to know that others also faced the same issue. Did you try something like archil or agentfs from turso?

We use QEMU for our sandbox infra. QEMU comes with qcow2 (copy-on-write), a native disk image format. It allows the VM to only write changes to disk which saves time and space specially if you want to snapshot and restore changes.

Wrote a piece on building a mental model for qcow2 overlays.

So it's Vue vs Next now?

Congrats on the launch! Legacy system users are also one of the slowest to adopt AI. How do you navigate that?

Who is using Mythos to find these things and where do they run it?

Building an execution layer for AI agents so people don't have to use Mac minis to run OpenClaw.

https://docs.celesto.ai

Building the execution layer for AI agents.

https://docs.celesto.ai

This is something solved by a bunch of new sandboxes including ours - SmolVM

https://github.com/CelestoAI/smolVM

you can use it with WSL or using our QEMU backend which supports Windows as well.

`SMOLVM_BACKEND="QEMU" smolvm pi start` should do the job.

I love using Pi with Codex and running multiple sessions on the same repo. To make my life simpler, added a single command to start a pre-installed Pi sandbox.

`smolvm pi start` is all you need!

Yeah totally relate to this. I’ve been talking to developers and engineers (~60 in total since last month). Most of them are just running YOLO without any security or safeguards while knowing that it’s dangerous.

One guy who works at a coding agent company just casually mentioned that we ask users consent that it can do damage and don’t apply any sandbox. Listening to this was mind boggling for me.

PS: we’re interviewing people as a part of user research for our sandbox product.

I've been building SmolVM -- an open-source AI sandbox focused on making it easy to run agents securely on laptops and on-prem servers.

Over last few weeks, we've improved speed, cold start, added pause/resume, host directory mounting, and a bunch of features.

You can use it with a CLI to manage sandboxes, run coding agents or use with Python SDK to build custom agent harnesses.

i was using neon and they had some similar feature but now using planetscale. would be curious to know how you all are doing it?

I've been dogfooding it by running coding agents. I mount my local filesystem and pull the changes back when done.

I've been working on an open-source isolated runtime for AI agents. Over a period of time I was able to bring the cold start time in sub-seconds, persistent filesystem, and network control.

I'm looking for feedback and if you're into this field - why or *why not* you would use it.

yes, we have an example of OpenClaw.

SmolVM uses microVMs which means the resource consumption is quite minimal. For now, you can specify the amount of RAM while starting the VM. We are making it elastic so you can configure to expand the RAM based on usage.

Here is the link to the [documentation](https://docs.celesto.ai/smolvm), but I can also show you on a call.

Also, I've been running OpenClaw using SmolVM on my macOS. It adds a hardware level isolation (using KVM).

https://github.com/CelestoAI/SmolVM/blob/main/examples/openc...

let me put some more in nvidia now

Also found this Vibe coding OSS from Modal - https://github.com/modal-labs/modal-vibe

Curious if Vercel is also gonna open-source theirs.

Nice! they forgot to add the Github link for the repo lol.

PS: If you're looking too - https://github.com/cloudflare/vibesdk

I’m using Neon too and upgraded to the scale up version today. Curious, what do you mean rhat they can get pricey?

You can run with LitServe. here is the code - https://lightning.ai/lightning-ai/studios/deploy-llama-3-2-v...

You can run it with LitServe (MPS GPU), here is the code - https://lightning.ai/lightning-ai/studios/deploy-llama-3-2-v...

LLM training in simple, pure C/CUDA. There is no need for 245MB of PyTorch or 107MB of cPython.